URLComponents.string should percent-encode colons in first path segment if needed

[jrflat]

The following code can recursively call .path(), causing an infinite loop.

let str = "./not a scheme:"
let baseStr = "base"
let base = URL(string: baseStr)!
let url = URL(string: str, relativeTo: base)!
print(url.path()) // Infinite loop

The issue is that self.path() calls .absoluteURL.path() if self has a base URL. .absoluteURL should never have a base URL, so this shouldn't recurse. However, if URL(string: absoluteString) fails like it does in this example, then .absoluteURL returns self, causing the recursion.

In the example above, merging the relative and base paths gives us ./not%20a%20scheme: since they are both relative. Then, we remove the dot segments as specified by the RFC algorithm, giving us not%20a%20scheme:. We assign this to a urlComponents.percentEncodedPath, then when we're finished building the components, we call urlComponents.string to get the absolute string.

URLComponents.string should always return a parsable URL string, or nil. In this case, the string that's returned is not%20a%20scheme:, which is not valid.

This PR fixes the issue by:

1. Percent-encoding colons in the first path segment when generating URLComponents.string, but only if the colons could be mistaken as a scheme separator.
2. Calling absoluteURL.relativePath() instead to ensure that we never recurse.

[jrflat]

@swift-ci please test

[itingliu]

For future reference: we prefer using XCTUnwrap to unwrap the string (like you did above on line 1327). I'd also appreciate if we can use XCTAssertEqual to assert the exact content instead of less-informational XCTAssertNotNil

[jrflat]

Ah that makes sense, I'll do a cleanup of all the URL tests in a follow-up PR to prefer that pattern